Question about muscle-only control for a reaching task using arm26.bioMod

[AleBulanti]

Hello,

I am currently working on a reaching task in Bioptim using the model "barm26.bioMod" and I need to know whether it is possible to use a model driven only by muscles, without residual joint torques. In my model, I modified the arm to include:

• 2 DoFs at the shoulder:

  • shoulder rotation around x
  • shoulder rotation around y

• 1 DoF at the elbow:

  • elbow rotation around x

My goal is to perform a reaching task using only muscle activations/excitations as controls.

I saw examples such as the static arm and muscle-driven examples, but I am unsure whether a pure muscle-driven formulation is feasible directly, or if residual torques are generally necessary for stability or convergence. Can you clarify this for me?

Thanks in advance!

[EveCharbie]

Hi,
Yes, you can have a pure muscle-driven simulation by specifying with_residual_torque=False

bio_model = MusclesBiorbdModel(biorbd_model_path, with_residual_torque=True)

However, residual torques are generally recommended for convergence because the muscle parameters (especially origin/via-points/insertion) are generally not perfectly defined. This may have an impact on the ability of the model to generate joint torques on the whole range of motion.

[pariterre]

Hi @AleBulanti!

I agree with everything @EveCharbie said. Since I strongly discourage avoiding residual torques, I would add that, if you absolutely need to remove them, I would recommend a two-pass optimization approach: first, run the optimization with residual torques enabled, and then perform a second pass without them.

Without residual torques, optimizers such as Ipopt and ACADOS can become trapped in inescapable local minima. This happens because, during the exploration of the solution space, the optimizer may enter regions where the muscles are effectively too weak or too strong to get out. Due to the gradient-based nature of these algorithms, they can get stuck there. Residual torques provide an alternative pathway by transferring some constraints to another linear system, which often helps the optimizer escape poor local minima and find better solutions.

That said, I still strongly discourage not using residual torques. Instead, it is often more useful to interpret them as a way to quantify the difficulty your model has in achieving a given task. If you strongly penalize residual torques in the objective function, the optimizer will only rely on them when truly necessary.